Econometric Methods for Panel Data

Transcription

1 Based on the books by Baltagi: Econometric Analysis of Panel Data and by Hsiao: Analysis of Panel Data Robert M. Kunst University of Vienna and Institute for Advanced Studies Vienna April 19, 2013

2 Outline Introduction Fixed effects Random effects Two-way panels Tests in panel models Poolability tests Testing for the presence of random effects The Hausman test Coefficients of determination in panels

3 Poolability tests The hypotheses of poolability tests These tests help select the panel model to be estimated, within the framework of fixed-effects models. For example: Are there individual effects or is it preferable to ignore them and to estimate by pooled OLS? Are there time effects on top of the individual effects? Are the coefficients β really constant across individuals? These tests follow a simple principle: restricted and unrestricted models are compared via LR or Wald statistics. Null distributions are F or χ 2, with degrees of freedom matching the number of restrictions.

4 Poolability tests Testing for individual effects Here, the null model (restricted) is y it = α+β X it +ν it, with iid errors, and the alternative (unrestricted) model is y it = α+β X it +µ i +ν it. The null hypothesis may be written as: H 0 : µ i = 0,i = 1,...,N.

5 Poolability tests Testing for individual effects: the statistic The traditional restriction-test statistic is F 1 way = (ESS R ESS U )/(N 1) ESS U /((T 1)N K), as there are N 1 restrictions from the maintained hypothesis to the null. There are (T 1)N K degrees of freedom in the unrestricted model. Under H 0, this statistic will be distributed as F N 1,N(T 1) K, assuming Gaussian errors.

6 Poolability tests Other analogous tests For the null of the pooled model and the alternative of the two-way model, the distribution of the F statistic will be distributed as F (N+T 2,NT N T+1 K) under the null, assuming Gaussian errors. For the null of the one-way individual-effects model and the alternative of the two-way model, the F statistic will be distributed as F (T 1,NT N T+1 K) under the null, assuming Gaussian errors. For large NT, χ 2 versions may be used instead, which are independent of Gaussian error assumptions.

7 Poolability tests Testing homogeneity of individual slopes In one specification, the unrestricted model is and the null can be expressed as y it = α+β i X it +µ i +ν it, H 0 : β 1 =... = β N = β,µ 1 =... = µ N = 0, with (K +1)(N 1) restrictions. The unrestricted model has only N(T K 1) degrees of freedom. If this test rejects, either the one-way model should be considered or individuals must be modelled separately. It may make more sense just to test the slopes, with K(N 1) restrictions.

8 Poolability tests Constellations in panel models and degrees of freedom coefficients β effects name β β i α OLS NT K 1 N(T K) 1 α i one-way N(T 1) K N(T K 1) α it two-way N(T 1) T +1 K N(T K 1) T +1

9 Testing for the presence of random effects Likelihood-ratio tests for variance parameters Assume L u is the likelihood of an unrestricted model, L r is the likelihood of a restricted model. A lemma says that the likelihood-ratio (LR) statistic LR = 2(logL u logl r ), will be, under the null of the restricted model, asymptotically distributed χ 2 with degrees of freedom matching the number of restrictions. The lemma requires several regularity conditions. In testing a variance parameter for zero, these conditions are violated and the property is not guaranteed to hold.

10 Testing for the presence of random effects LR test for random effects in two-way panels The unrestricted RE model y it = α+β X it +u it, u it = µ i +λ t +ν it, with the restricted null H 0 : σµ 2 = σ2 λ = 0 violates the regularity conditions. The LR test statistic has the non-standard distribution 1 4 χ2 (0)+ 1 2 χ2 (1)+ 1 4 χ2 (2), where χ 2 (0) denotes point mass at zero.

11 Testing for the presence of random effects LR test for random effects in one-way panels The unrestricted RE model y it = α+β X it +u it, u it = µ i +ν it, with the restricted null H 0 : σµ 2 = 0 again violates the regularity conditions. The LR test statistic has the non-standard distribution 1 2 χ2 (0)+ 1 2 χ2 (1). An analogous test can be used for random time effects. These tests are due to Gourieroux, Holly & Monfort.

12 Testing for the presence of random effects LM tests for random effects Lagrange-multiplier (LM) tests have standard χ 2 asymptotics. They have usually less power. One may test for two-way random effects using pooled-ols residuals ũ by LM: LM = LM 1 +LM 2, LM 1 = LM 2 = NT 2(T 1) NT 2(N 1) { 1 ũ (I N J T )ũ ũ ũ { 1 ũ (J N I T )ũ ũ ũ } 2, } 2, or one may test for one-way effects by using LM 1 or LM 2. This test is due to Breusch & Pagan.

13 The Hausman test The Hausman test principle Hausman tests can be used in all situations where two model specifications and two estimators are available with the following properties: 1. In the restricted model (null), the estimator ˆθ is efficient, the estimator θ is consistent though typically not efficient; 2. in the unrestricted model (alternative), the estimator ˆθ is inconsistent, the estimator θ is consistent. Then, the difference q = ˆθ θ should diverge under the alternative and it should converge to zero under the null. Moreover, under the null q and ˆθ should be uncorrelated.

14 The Hausman test The RE and the FE model The null of the RE and the alternative of the FE model correspond to the Hausman situation: 1. In the RE model, the GLS-type RE estimator is efficient by construction for Gaussian errors, the FE estimator and even the OLS estimator are consistent; 2. in the FE model, the RE estimator is inconsistent, because of the omitted-variable effect, while FE is consistent by construction.

15 The Hausman test Two views on the RE inconsistency in the FE model 1. The estimator (X Ω 1 X) 1 X Ω 1 y is inconsistent, if the true model is y = Xβ+Z µ µ+ν, as the regressors Z µ are omitted; 2. Mundlak showed that the RE estimator for a stochastic regression model with X and Z µ correlated is identical to the FE estimator: the RE estimator imposes independence of effects and covariates. Some argue, however, that Mundlak s alternative is not really the same concept as the fixed-effects model.

16 The Hausman test The Hausman test statistic The Hausman test statistic is defined as m = q (varˆβ FE varˆβ RE ) 1 q, with q = ˆβ FE ˆβ RE. Under RE, the matrix difference in brackets is positive, as the RE estimator is efficient and any other estimator has a larger variance. The statistic m is distributed χ 2 under the null of RE, with degrees of freedom determined by the dimension of β, K.

17 R 2 in a panel model Should the variation due to effects be part of the explained variation or not? If yes, the R 2 has little to say on the importance of the substantial covariates β. There is no unanimous agreement on which R 2 to report in a panel. Some programs (STATA etc.) follow the suggestion by Wooldridge and report three measures: within R 2, between R 2, and overall R 2.

18 The within R 2 For the within R 2, the total sum of squares TSS is assumed as TSS = N T (y it ȳ i ) 2, i=1 t=1 i.e. after adjusting for effects. Then, also residuals in the residual sum of squares ESS rely on the purged model, and the statistic 1 ESS TSS is maximized by LSDV or the FE estimator.

19 The between R 2 For the between R 2, the total sum of squares TSS is assumed as TSS = N (ȳ i ȳ) 2, i=1 i.e. just N individual time averages. Residuals in the residual sum of squares ESS also rely on the between model with N observations, and the statistic 1 ESS TSS is maximized by the between estimator.

20 The overall R 2 Finally, the overall R 2 relies on the usual TSS definition TSS = T t=1 i=1 N (y it ȳ) 2, and on residuals calculated without accounting for effects. It is maximized by pooled OLS. Values for the three R 2 may be compared. For example, if within and overall R 2 are close, this is evidence for individual effects being not so important etc.